What Quasar AI is
Quasar AI is the project's own intelligent system. Its C++ core, knowledge, language analysis and inference are developed within the project and run on our own infrastructure.
In supported scenarios, answers are built by algorithms over knowledge from books, dictionaries and manuals. Each case preserves its grounds, assumptions and revision history. When evidence is missing, the system should name the gap. This is a development principle; universal understanding of arbitrary requests has not yet been achieved.
What has been verified
In the mathematical scenario, Quasar calculates a 2×2 matrix determinant exactly, compares two input editions and independently checks the difference between their results. In the control example, changing one cell from 4 to 5 changes the result from −2 to −1.
Withdrawing an assumption removes the applicability of dependent conclusions while retaining the earlier result in the history. Cases survive a restart; a person can select a particular answer and continue that case.
The same mathematical meaning can already be expressed in several orders of presentation. A first Russian construction is built from roles — “Детерминант равен −2”: the designation has supporting evidence, the word form comes from morphology, and an independent check verifies agreement and meaning.
The QLEX dictionary package is connected to the standard engine. For tested dictionary queries, answers are built from typed articles with an internal trace and source coordinates; unsupported meanings are not replaced by guesses.
For ordinary ice in fresh liquid water near 0 °C, Quasar has checked an inference from density values and Archimedes’ principle. The original question and five paraphrases produce the same ten-step trace; that trace is not reused for a different temperature or medium without supporting evidence.
These results apply to the specified scenarios. They do not yet amount to a general conversational assistant, translator or autonomous programmer.
What we are working on now
Updated 25 September 2026.
The project already contains large Wiki and Habr knowledge cubes, dictionaries, physics knowledge and programming manuals. The current question is not whether to add more gigabytes, but whether Quasar can complete the whole path: understand a request, select a relevant source, retrieve a specific record, connect it to a rule and preserve the grounds for its answer. A large cube does not by itself prove that Quasar can use its knowledge.
When a request is ambiguous, Quasar should first check available articles and other relevant sources, then compare them with the history of that conversation. If the sources and context identify one meaning, it should select it and preserve its grounds. A clarifying question is asked only if distinct, source-supported possibilities remain after that search or a specific fact needed for retrieval is missing; it should resolve that distinction rather than replace knowledge retrieval with a refusal.
First, we are testing whether Quasar can use the contents of the cubes already present; new data will be added only after a specific knowledge gap is demonstrated. Then we will extend reusable inference from facts and rules with explicit applicability conditions. The next major goal is constructing code from language manuals: derive program constructs and lines from documented rules, rather than calling a code generator or inserting ready-made solutions or task templates. A capability not confirmed by the language specification or manual must not be presented as supported. New language features are a separate development effort or a proposal to the language maintainers.
The presence of 19 languages and 45 directions in the dictionary corpus does not mean full-text translation is ready for all of them; open-ended conversation and autonomous programming have not yet been achieved.
The laws it works by
Freedom to research anything, with a ban on presenting the unproven as proven. An honest “not proven” that names the missing observation is preferable to a confident assumption.
Provenance is never fabricated: an unknown source remains unknown in the journal. A feature is evidence only where it distinguishes the cases under consideration.
Accepted milestones are protected by targeted tests and negative checks. The full integration gate remains a separate upcoming milestone.
What this has to do with the trading terminal
The ecosystem's released products work independently. Applying the research core to trade reviews and explanations of risk remains a development goal.
Quasar AI has not yet been released as a user-facing product. Research continues through verifiable milestones, without a promised completion date.